ML Engineering & Deep Learning
Applied ML and theory
One ML system that operators run every day, and the theory that tells you how hard you can push a training run. Both come with numbers you can check. Open a project for the full case study.
-
Forecasting · deployed
Telecom sites in Cameroon were losing diesel and nobody could see where. This predicts what each site should burn and flags the clusters and sites whose predicted use stands out. Operators still run it.
NSE 0.982 · 84,617 liters of fuel accounted for
A gradient-boosting regressor behind a Flask web app, with a deviation threshold at the mean plus 2 standard deviations, a monitoring dashboard, and exports for audits.
-
Doctoral research · published
Analysis of training neural networks
Springer Nature 2024How large a training step you can take before a model stops learning, worked out in advance instead of found by watching a run fail.
The maximum learning rate does not decay exponentially with depth
Global convergence of gradient descent for deep linear networks, carrying earlier gradient flow analyses into the discrete setting and proving convergence to a global minimum for almost all initializations in the two-layer case.
-
Research · preprint
Convergence of SGD in deep linear networks
arXiv 2026A proof that training on random mini-batches, as done in practice, keeps a deep linear network's weights bounded and drives its gradient to zero, with step-size rules you can compute.
Almost sure convergence of SGD to critical points, with no Lipschitz-gradient assumption
Extends gradient descent convergence results for deep linear networks to mini-batch SGD on a regularized square loss. It proves the iterates stay bounded and, through asymptotic pseudotrajectories from stochastic approximation, that SGD converges to the critical set under deterministic and adaptive step sizes.